Compaction fails: thinking/redacted_thinking blocks in latest assistant message cannot be modified #41738

Summary

Compacting a session that uses an extended-thinking model fails with a Claude API validation error. The compaction request is rejected before it can run, so the session can no longer be compacted.

Error

undefined: The model returned the following errors: messages.49.content.13: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

(The message/content indices vary per session.)

What I think is happening

When extended thinking is enabled, assistant messages contain thinking / redacted_thinking content blocks that Anthropic signs. On resend they must be byte-for-byte identical to the original response — they cannot be edited, truncated, reordered, or dropped.

Compaction rewrites/summarizes the conversation history and appears to mutate or partially copy a thinking block in an existing assistant message (rather than passing it through verbatim). The API then rejects the request with the error above.

Steps to reproduce

1. Use a Claude model with extended thinking / reasoning enabled.
2. Have a conversation long enough that at least one assistant turn contains a thinking block.
3. Trigger compaction.
4. Compaction fails with the error above and cannot complete.

Expected

Compaction should preserve thinking / redacted_thinking blocks verbatim (or drop whole assistant turns cleanly) so the resulting request passes Claude's validation.

Environment

• OpenCode version: 1.18.4
• Model: Claude (extended thinking / reasoning enabled)
• Platform: Linux

Workarounds observed

• Starting a fresh session avoids the corrupted history.
• Disabling extended thinking for the model lets compaction proceed (no signed blocks in history).